Controlling the Output Format

Reviewed & published by Brayan K

Telling the AI exactly how to shape its answer — as a table, JSON, bullet points, or a one-paragraph summary — gets you a result you can use immediately, instead of a wall of prose you have to reformat by hand.

Part of the free Prompt Engineering course at LearnCodingFast — hands-on lessons with worked examples and the output they print, plus practice exercises and a quick quiz.

You almost always know where the answer is going next: a spreadsheet, a program, a slide, a quick scan. Say so. This lesson shows you the formats worth asking for and how to nail length, too.

What You'll Learn in This Lesson

• Why naming a format saves you re-prompting
• Asking for tables and CSV for spreadsheet-ready data
• Getting clean, parseable JSON for programs
• Bullet lists, numbered steps, and markdown
• Controlling length: word limits, bullet counts, one sentence
• Forbidding preamble: 'just the answer, nothing else'

1 Say where the answer is going

If you don’t name a format, the model defaults to paragraphs — and you waste a round re-asking. Same question, two prompts:

Tell me about the 5 biggest
planets in the solar system.

You get rambling prose you’d have to reformat to use anywhere.

List the 5 biggest planets as a table
with columns: Planet, Diameter (km),
Moons. Sort largest first. Table only.

Spreadsheet-ready, sorted, no fluff — usable the instant it arrives.

2 A catalog of formats to ask for

For JSON, add “return ONLY valid JSON, nothing before or after” so a chatty intro doesn’t break your parser.

3 Control the length precisely

“Keep it short” is vague. Concrete limits work far better:

• "in exactly 3 bullet points"
• "under 100 words"
• "one sentence"
• "a single short paragraph"
• "the answer only, no introduction or sign-off"

📋 Copy-paste output-format template

<your request>.
Format the answer as <table / JSON / bullets / etc.>.
<For JSON: with keys: ...; return ONLY JSON.>
Length: <3 bullets / under 100 words / one sentence>.
Do not add any introduction or sign-off.

Try it for real

Asking for JSON does not get you JSON — it gets you something that usually looks like JSON. Run this on three real-shaped replies: one clean, one wrapped in chatter, one with a trailing comma a human reads straight past. The prompt asks for the format; the parser is what makes your program safe when the reply drifts.

# The rule this lesson is really teaching: never trust a format you did not verify.
# Below are three replies a model gave when asked for JSON. Two of them look fine.

import json

replies = [
    '{"name": "Ada", "role": "engineer"}',
    'Sure! Here is the JSON:\n{"name": "Ada", "role": "engineer"}',
    '{"name": "Ada", "role": "engineer",}',
]

REQUIRED = {"name", "role"}

for i, raw in enumerate(replies, 1):
    print(f"--- reply {i}: {raw!r}")
    try:
        data = json.loads(raw)
    except json.JSONDecodeError as e:
        print(f"    REJECTED — not parseable: {e.msg} at column {e.colno}")
        continue
    missing = REQUIRED - data.keys()
    if missing:
        print(f"    REJECTED — missing keys: {sorted(missing)}")
    else:
        print(f"    ACCEPTED — {data}")

print()
print("Reply 2 is valid JSON wrapped in chat. Reply 3 has a trailing comma, which")
print("a human reads straight past. Both are why 'reply in JSON' belongs in the")
print("prompt AND a parser belongs in your code.")

# 🎯 YOUR TURN — write the guard that decides what your program accepts.

import json

REQUIRED = {"title", "priority"}

replies = [
    '{"title": "Fix login", "priority": "high"}',
    '{"title": "Fix login"}',
    'here you go: {"title": "Fix login", "priority": "low"}',
]

def validate(raw):
    # 1) Anything that is not parseable JSON is rejected outright.
    try:
        data = json.loads(raw)
    except json.JSONDecodeError:
        return "not JSON"

    # 2) A parsed object still has to carry every required key.
    missing = REQUIRED ___ data.keys()
    # 👉 replace ___ with  -

    return "ok" if not missing else f"missing {sorted(missing)}"

for raw in replies:
    print(validate(raw))

# ✅ Expected output:
# ok
# missing ['priority']
# not JSON

Common Mistakes (and the fix)

• Not naming a format. — You get prose; say “as a table” or “as JSON”.
• Asking for JSON without “only JSON”. — A friendly sentence sneaks in and breaks parsers.
• Saying “keep it short”. — Give a number: 3 bullets, under 100 words.
• Forgetting to name the columns/keys. — Specify them so the structure is predictable.

⏱ Test Yourself — Timed Quiz

10 quick questions, 12 seconds each. Instant feedback — beat the clock!

🎉 Lesson Complete

• ✅ Name the format — table, JSON, bullets — so the answer is usable now
• ✅ For JSON, demand “only JSON, nothing else” and name the keys
• ✅ Control length with concrete limits, not “keep it short”
• ✅ Forbid preamble when you want just the answer
• ✅ Picture where the output lands and ask for that exact shape

Practice quiz

Continue this course

• Previous: Chain-of-Thought Reasoning
• Next: Prompting AI to Write Code — Give language, version, the error and the goal — then read and test the result

Frequently asked questions

Why should I tell the AI what format I want?

Because you almost always have a use in mind — a spreadsheet, a slide, code, a quick scan. If you don't say, the model defaults to prose, and you waste a round re-asking for a table or list. Naming the format up front ('reply as a 3-column table') gets you something usable on the first try.

How do I get clean JSON I can use in a program?

Ask for it explicitly and name the keys: 'Return ONLY valid JSON with keys name, price, and inStock. No text before or after.' The 'nothing before or after' part matters — otherwise models often wrap the JSON in a friendly sentence that breaks parsers.

Can I control how long the answer is?

Yes, and it's one of the most useful controls. Be specific: 'in exactly 3 bullet points', 'under 100 words', 'one sentence', or 'a single paragraph'. Vague asks like 'keep it short' work less reliably than a concrete number.

What formats can I ask for?

Tables, CSV, JSON, bullet or numbered lists, markdown with headings, a single paragraph, a checklist, or 'just the answer with no preamble'. Pick whatever matches where the output is going next.